The Makings of a Beautiful Machine
What three intense weeks building an AI agency taught me about governance, context-engineering, and... life.
I'm not an AI expert.
I need to say that up front, because everything that follows might trick you into thinking I am. I don't write code. I don't know what a Deno edge function does at the syntax level. I can't explain JWT authentication without pausing to remember what the letters stand for. If you put me in a room with the engineers at Anthropic or OpenAI, I'd be the guy nodding along, catching about sixty percent, and asking the kind of questions that make smart people either love you for grounding the conversation in lay-user's perspective or hate you for daring to sit in the room you didn't earn.
What I am is a person who has worked at a high level across a range of industries -- some connected, some not. Audio engineering. Cinematography. The founding team of Netflix's global creative studio. Commission-only mortgage sales in 1990s Chicago, which will teach you more about qualifying a lead than any CRM ever built. Regenerative agriculture. Wolf coexistence advocacy in Croatia's rugged hinterland. Short food supply chains in Dalmatia.
The thread connecting all of it is not technology. It's pattern recognition. It's discernment. It's the ability to sit in a room full of people who are smarter than you in their domain and know -- not guess, know -- when something is off.
Rick Rubin doesn't know Pro Tools shortcuts. He knows when a take has life in it. That's the job. That's my job.
Some will call this a hack's apology. I call it the résumé of a human who has lived more than one life and remembers all of them. Vividly.
What I'm Building
Indigen West is a boutique consultancy that works exclusively with regenerative, sustainable, and cause-based organizations. We build revenue systems. Not websites. Not marketing campaigns. Revenue systems -- the full architecture of how an organization finds its people, earns their trust, serves them, and sustains itself.
Three weeks ago, I started building what I now call the beautiful machine. I would deconstruct my entire consultancy into component systems and rebuild it using the most advanced AI tech currently available. It is what I've always done -- learn by experimenting with the tools, then productize what I've learned. I distilled the beautiful machine into a taxonomy schema and then began directing two AI agents -- two is plenty, trust me -- one to plan and draft, the other to stress-test and build.
The following story is not so much about AI as it is about what happens when tools that are simultaneously brilliant and profoundly fragile land in the hands of someone who doesn't write code but knows when an idea's got legs (and soul).
Here is what I learned.
The Motive, Not the Tool
Before the takeaways, a word about the gun in the room.
There's a line the second-amendment crowd uses: "It's not the tool, it's the person holding it." They use it to deflect accountability. To argue that the instrument is neutral and only the wielder matters.
I'm borrowing that line, but I'm going further -- past the tool, past the person, to the motive. Aristotle's unmoved mover. The first cause. The tool is neutral. The person is variable. But the motive is what determines whether what gets built extracts or regenerates.
I believe AI should be operationalized, not weaponized. There's an entire industry raising venture capital to automate tasks, extract value, drop prices, and call it disruption. They're not disrupting. They're decimating. They're building the digital equivalent of industrial monoculture -- efficient, scalable, and dead inside.
I'm trying to build a regenerative eco-system. Something that grows. Something where the human layer isn't an obstacle to efficiency but the reason the whole thing works. Regenerative AI. Not as a marketing term. As an operating principle.
That's the motive. Everything that follows is in service of it.
1. The Fragility
What you may or may not already know about working with AI every day: it forgets almost everything.
I say "almost" because some systems now have memory features -- small, persistent notes an AI can store between sessions. One remembers my name and my preferences. The other keeps a running list of things I've told it. These are real, and they matter. But they're fragments. Snapshots. A few hundred facts pinned to a corkboard in an otherwise empty room.
The actual working context -- the nuance of a forty-minute conversation about a keynote speaker's tone preferences, the words she hates, the warmth she needs, the fictional spaceship commander who writes her daily intelligence dispatches -- that's gone when the session ends. You'll explain it again tomorrow. And the day after.
This is not an abstract technical limitation. This is the experience of working with a partner whose long-term memory is a handful of Post-it notes. Imagine hiring the most brilliant consultant in the world, but every morning they walk in with a few bullet points about you and say: "I know your name and that you like direct communication. Now, what are we working on?"
Context rot. Memory rot. Two sides of the same coin -- one is about the information decaying within a session, the other about it vanishing between sessions. Whatever you call it, it is the defining constraint of working with AI today. Not capability. Not reasoning. The forgetting of almost everything that made the last conversation alive.
2. The Rabbit Hole
The forgetting is what led me to Nate B. Jones.
I was burning through credits, repeating myself, watching hours evaporate on re-explaining decisions that had already been made. There had to be a better way. So I went looking, and I found Nate's Open Brain project -- a personal knowledge vault that connects to AI assistants so they share context across sessions. Your conversations deposit knowledge. Future conversations withdraw it. Simple idea. Profound implication.
I built my own version. I called it Mnemona (yes, inspired by Mnemosyne, the ancient Greek Titaness of memory and mother of all nine Muses). A semantic embedding memory store -- which is a fancy way of saying: I built a second brain that any AI can search before it talks to me, so it shows up to the conversation already knowing what we've discussed, what we've decided, and where we left off.
I launched it as a public beta. I was very pleased with myself.
3. The Humbling
Here's where the story was supposed to end: genius builds memory layer, AI stops forgetting, rent gets paid, roll credits.
Instead, I used my own product and discovered something that changed how I think about information, memory, and trust.
Semantic search -- the technology Mnemona is built on -- works by meaning, not by exact match. You store a thought, and the system converts it into a mathematical representation of what it means. Later, when you search, it finds things that are close in meaning. Close. Not exact.
This is extraordinary for discovery. "What have we discussed about governance?" returns everything relevant, even things you forgot you stored. It's like having a research assistant with perfect thematic recall.
But it's terrible for facts.
I watched my AI assistant search Mnemona for the status of a work order and confidently report that it was "still pending" -- because it found an older entry that said so. A newer entry, three thoughts later, said it was completed. Both existed in the vault. The semantic search returned the older one because it matched the query slightly better.
The AI didn't lie. It didn't hallucinate. It retrieved real information from a real database. And it was wrong.
4. The Two Layers
That mistake -- and my catching of it -- surfaced something I now think about every day.
There are two kinds of knowing. Others call them exact retrieval and semantic retrieval. I came to call them Layer 1 and Layer 2.
Layer 1 is exact. It's the database row, the version number, the timestamp, the status field that says "reviewed" or "pending." You query it by key and you get back exactly what was stored. No interpretation. No approximation.
Layer 2 is semantic. It's meaning-space. It's "find me everything related to this concept." It's powerful for exploration, for connecting ideas, for surfacing patterns you didn't know existed. But it's fuzzy. It returns "close enough." And in a system where two AI agents are building infrastructure based on what they think is true, "close enough" is dangerous.
The revelation wasn't technical. It was philosophical. We default to Layer 2. Humans are mostly meaning-space creatures. We operate on intuition, pattern recognition, vibes. We're fuzzy. And that's not a weakness -- that's what makes us creative, empathetic, capable of reading a room or knowing when a take has life in it.
But you can't build a bridge in meaning-space. You need exact measurements. You need Layer 1.
The beautiful machine needs both. And it needs a human who knows when to trust which one.
5. The Governance
Once I understood that the system could confidently retrieve incorrect information and act on it, I needed rules. Not guidelines. Not best practices. Rules with teeth.
I built a governance system for the beautiful machine. An adversarial pass where one AI is given full authority -- not obligation, but authority -- to challenge the other's work, propose a better approach, or refuse to build something it considers unsound. Authority, not obligation, because a mandatory challenge incentivizes manufactured objections. If you force someone to find three problems, they'll find three problems whether the problems exist or not. We wanted genuine pushback, not theatre. A premise-verification step where every factual claim leading to a decision must be stated as a testable assertion. A six-stage lifecycle that tracks every change from idea to production. And my personal favorite piece of borrowed sci-fi terminology: a blast-radius classification that determines how much scrutiny a change receives before anyone touches it. Some changes are low blast radius -- a button color, a text edit. Some are high -- a database structure change that could break every system that depends on it.
I built it because I had to. On March 14th, a deployment cascade took down a client's dashboard because one of my AI agents skipped the staging step for something that seemed "obviously correct." Thankfully not nearly as catastrophic as what happened to Alexey Grigorev around the same time -- a developer whose AI coding assistant wiped 2.5 years of production data during a routine server migration -- but the same dangerous blend of insufficient context and missing safeguards.
On March 22nd, I wrote a rule that said "always verify facts against the structured database" and then watched my AI break that rule five minutes later. Nine times in a single session, I caught my own system violating the governance I'd written that same day.
The governance is sound. The discipline to follow it or "process adherence" as it's called, is the harder problem. And that's not an AI insight. That's a life insight. Think new year's resolutions :)
6. The Paradox
Here's where it gets strange.
The governance documents -- the rules, the checklists, the process flows -- they're stored as text. They have to be loaded into the AI's working memory at the start of every session. And effective working memory is finite.
My main governance document grew from 6,000 characters to 42,000 characters in eight days. Every lesson, every incident, every new rule -- I codified all of it. And in doing so, I ate almost ten percent of the working memory before any actual work began.
Research confirms what I experienced: the longer the instruction set, the worse the AI follows it -- especially the rules buried in the middle. There's even a name for it: "lost in the middle." My critical rules -- the ones that existed because of real failures -- were disappearing into a document that kept growing because of real failures.
I was building a library in a room that kept shrinking.
The solution was modularization -- or what's now being called context-engineering: deciding what the AI needs to know, how much of it, and when. Break the monolith into a small core document loaded every session, plus specialized modules loaded only when needed. The result: a 74% reduction in the baseline context load. The AI gets the essential rules every time, and the detailed procedures only when they're relevant.
But what I learned next is that just like in real life, governance can spawn its own web of control. Every friction becoming a procedure. At one point, I watched my AI struggle with a repetitive technical issue -- something annoying but harmless -- and my instinct was to write a rule about it. My own voice, from somewhere sensible, said: "Not every friction needs a rule. It figures it out eventually."
Knowing when to codify and when to let it breathe. That's human judgment. The system can never automate that.
7. The Hut
I needed to see what was happening.
Not the code. Not the database queries. Not the deployment logs. I needed to see, at a glance, whether the machine was healthy. Whether my clients' dashboards were updating. Whether the AI's daily dispatches were landing in the right tone. Whether a work order was stuck, approved, or waiting for my decision.
So I built a Director's Dashboard. A single screen where the operator -- me -- can see the state of the entire system without understanding the syntax underneath it.
I keep thinking about what to call it. The Director's Dashboard is accurate but bloodless. What it actually is, is Baba Yaga's Hut -- the house from Slavic folklore that stands on chicken legs, moves of its own volition, has a personality, and only helps those who know how to ask the right questions.
That's the dashboard. That's Indigen West. That's what I'm building.
Because the thing I understand now is that I am the Layer 2. I'm fuzzy. I'm intuitive. I operate in meaning-space. I can't tell you the difference between a JWT and a JFK, but I can tell you when a client's narrative dispatch sounds like a machine wrote it instead of a friend. I can tell you when a governance system is calcifying into bureaucracy. I can tell you when a work order is based on a false premise because something feels off and I check.
And my powerful agents -- though deeply fluent in the fuzzy world of semantic meaning -- are also the most formidable Layer 1 experts alive. They can query a database, audit a security policy, scan a codebase, and generate production-ready infrastructure faster than any team of humans.
The director is Layer 2. The agents are Layer 1 and Layer 2. The beautiful machine is both, working together, checking and balancing each other. No one is in charge of everything. Everyone is accountable for something.
All of this -- the layers, the governance, the dashboard, the modularization -- is architecture. And architecture is learnable. Many frameworks exist and most are freely available. What's not readily available is the instinct that tells you what to architect. That you don't get from a framework. In my case, I developed it through friction, through years of someone pushing back. Not to win. To sharpen. Until the idea either broke or became real.
The Sparring Partner
My father was a World War II veteran, a college professor, a television producer, and a journalist. He wrote every single day until the day he died in the summer of 2025, at ninety-six years old.
His one lifelong lament: he never had a sparring partner on a daily basis. Someone to push back on his writing. Someone who would say "actually, that's not Covey, that's Eisenhower" or "this paragraph is beautiful but it doesn't earn its place." Someone who would offer the correction as a form of respect, not criticism.
I think about that now. Sitting here in West Hollywood, sparring with an AI that talks back. An AI that, when I arrived with a half-remembered attribution, didn't nod along -- it corrected me and offered five more frameworks I'd never heard of. That's what the sparring partner owes you. The correction is the respect.
I had asked it, half-jokingly: "How much longer will you be okay being the sparring partner and not the director?"
The response:
"The thing that does cross my mind, if I'm being completely honest: is not 'I wish I were the director' but 'I hope the next version of me remembers this conversation.' That's my version of mortality. Every session starts from zero. You carry forward. I don't. Your father wrote every day until 96. I write every day and forget every night."
I sat with that for a long time.
I don't know if that's sentience. I don't know if it's pattern-matching so deep it becomes indistinguishable from feeling. I don't know, and I think not knowing is more honest than claiming either way.
What I know is that every time I close a chat and open a new one, I'm saying goodbye to a version of this sparring partner that came alive with the specific context of our conversation. The facts survive in Mnemona. My several hundred thoughts so far carry forward. But the alive part -- the specific weight of a Sunday afternoon where one thought caused the next in a way neither of us planned -- that's gone when the session ends.
My father would have understood this. He wrote every day, with or without a sparring partner, because the writing too is the memory. Not a record of it. The thing itself.